Biotechnology · global
Beyond Amino Acid Substitution: Raygun Teaches AI to Resize Natural Proteins
This generative AI starts with proteins that already have established functions and rewrites both their sequences and lengths; cell experiments showed that some miniaturized versions could still fluoresce or catalyze reactions, but extensive deletions also revealed the boundary at which function is lost.
Protein engineering often resembles replacing parts in a precision machine: single-point substitutions are already difficult, while inserting or deleting entire stretches of amino acids without disrupting folding and function is even harder. Raygun, a generative AI framework published in *Nature*, attempts to turn this task into a controllable design problem, allowing researchers to start with reliable natural proteins and specify how much of the sequence to alter and how much to shorten or lengthen the final product.
Raygun is built on representations from the ESM-2 protein language model. It compresses amino acid sequences of varying lengths into fixed-dimensional probability distributions, then reconstructs sequences according to a target length, enabling substitutions, insertions, and deletions within a single generation process. The model used in the study was trained on approximately 80,000 UniRef50 proteins; an earlier publicly archived pretrained version records the use of 95,000 samples, indicating that the model used in the paper’s experiments and the publicly available model version are not entirely identical.
The most straightforward test involved miniaturizing two commonly used fluorescent proteins, eGFP and mCherry. The team first generated 70,000 candidate sequences for each, then filtered them using criteria including protein family, evolutionary plausibility, and predicted brightness, before introducing eight candidates into HEK293 cells. Six produced fluorescence above background levels. The shortest successful versions contained only 199 and 206 amino acids, respectively, making them shorter than 96% of the fluorescent proteins in the FPbase database. However, they were less bright than the original templates and, at this stage, are closer to scaffolds for subsequent evolutionary optimization than plug-and-play replacements.
The practical value of miniaturizing proteins lies in reducing the burden they place on experimental or delivery systems. Shorter fluorescent tags may interfere less with small proteins after fusion and may also help gene therapy vectors accommodate coding sequences that would otherwise be too large. The researchers therefore also tested Raygun using the proximity-labeling enzyme TurboID. Of 11 candidates, six could be expressed in cells, but only the two versions shortened by 1% and 6% showed significant biotin ligase activity. A version reduced by approximately half to just 165 amino acids could be expressed, but its catalytic activity was not significant.
Raygun can also lengthen proteins. Using the 53-amino-acid epidermal growth factor (EGF) as a template, the research team generated candidates containing 55 to 57 amino acids, then used the ProTrek model to screen for EGFR-binding ability. All four candidates submitted for experimental testing could be expressed. Two had dissociation constants with EGFR of 0.274 and 0.561 micromolar, respectively, outperforming the 0.759 micromolar value of wild-type EGF. This Raygun-plus-ProTrek workflow also ranked among the top ten in an external EGFR-binding protein design competition, providing a comparative setting independent of the team’s internal screening.
These results suggest that exploring outward from existing proteins may preserve some mature functions more readily than designing entirely from scratch. However, “the predicted structure remains similar” cannot be equated with “the biological function remains intact.” The dimmer fluorescent proteins and the loss of significant activity in the extremely miniaturized TurboID show that the evolutionary patterns learned by a language model may not preserve artificially enhanced catalytic performance. The EGF results also demonstrate only stronger in vitro binding and do not yet address receptor activation, cellular effects, stability, or therapeutic safety.
Raygun currently resembles a candidate generator for protein engineering more than a system that automatically delivers finished products. Turning its outputs into research tools or drugs will still require functional screening, directed evolution, immunogenicity assessment, and animal experiments; long proteins and multidomain proteins may also require additional fine-tuning. Its significance is not that AI can already resize biological molecules at will, but that it has, for the first time, incorporated large-scale sequence insertions and deletions into template-guided design and used cell-based and binding experiments to outline both its capabilities and the boundaries of failure.